What I've found for real-world shooting, is that ETTR works best with lower-contrast scenes where you don't really have any whites. For scenes where there are highlight tones that I care about maintaining discernible detail, I get better results by placing them 1/3 to 1/2 stop below clipping, rather than right under the clipping limit. Even if that means having to add a small boost to the shadows or darker midtones, noise is pretty much a non-issue at base ISO and this approach gives me more natural-looking results with less fiddling around in the raw converter.
Since the original post by Michael on ETTR, it has become almost a religion for some photographers. However, the originally stated rationale for ETTR concerning the number of levels in the brightest f/stop of the image is incorrect and the naming of the process is based on the appearance of the the histogram, and the appearance of the histogram can be misleading. The basic concept of ETTR is to give as much exposure as practicable in order to collect the maximum number of photons, therey increasing the signal to noise ratio and maximizing dynamic range. An image taken at high ISO with the histogram fully to the right may have captured fewer photons than an image taken at a lower ISO and showing a histogram fully to the right. These principles are explained in great detail by Emil Martinec
Some purists (such as Panopeeper) allow no headroom and place the highlights at clipping, while Jeff and others allow some headroom for protection of the highlights. If the sensor is linear up to clipping, headroom should not improve the image but can help prevent inadvertent clipping of the highlights. The advangtages of ETTR can be overblown. Since signal to noise varies as the square root of exposure, doubling of the exposure will improve S:N only by a factor of 1.4. With current sensors, this may not make much difference.
Maximizing the number of photons collected requires the use of base ISO, so the principles of ETTR can only be fully realized at base ISO. If shutter speed/aperture considerations prevent fully exposing to the right at base ISO, then one can use a higher ISO. Under these conditions, shadow S:N will be improved only to a certain point by increasing the ISO and having a histogram with data on the right. Above a certain ISO (often 1600 on many cameras), the histogram will look better with a higher ISO but the signal to noise ratio in the shadows will not be better than would be obtained at ISO 1600 with a histogram to the left. At this point, it is best to increase exposure in the raw converter. The higher ISO under these conditions will limit headroom and dynamic range. This is explained in detail in Emil's paper.
While these concepts are beyond the beginner lever, the take home message is to give as much exposure (shutter speed and f/stop) as possible. Signal to noise is largely deterined by exposure. A high ISO image will have more noise because of less exposure rather than the use of a high ISO per se.